Evidence against and open controversies

This file gathers the evidence against the figure rules: null results, disputed studies, and advice that a study contradicts, by theme.

Minimalism against embellishment

Rule under test: remove ink that is not data (Tufte, as reported by Few 2011 and Europa 2023); drop decoration (Rougier 2014).
Bateman 2010: 20 students; embellished charts were described as accurately and, with 10 participants in the long-term condition, recalled better after two to three weeks. Grade: measured, for the charts of one designer.
- The embellishments came from one illustrator, and exact values were not read (Kosara 2010); Few 2011 adds that the messages fit in one sentence and the embellishment was relevant to them.
- Few 2011 restates the result with those limits and criticises the study. Kosara 2010 agrees minimal charts still suit analysis.
Borgo 2012 (abstract): recall up, visual search slower.
Skau 2015 (abstract and Kosara's summary): most bar embellishments lowered accuracy when comparing bars.
Cheng 2026 (held): interesting irrelevant material lowers learning, g = -0.16.
Reading: decoration that carries the message may help memory; decoration that distorts a quantity or distracts hurts. The rule "remove all decoration" is too broad and "add some" is unsupported for diagrams.

Inbar 2007 (abstract): 87 students preferred a standard bar graph to a minimalist one. Hill 2018 (secondary, from a search summary): charts with lower data-ink ratios were rated better.
Both measure ratings. Skau 2015 measures accuracy and finds some embellishments cost it.
Reading: do not take a rating as a reading result.

Burch 2021 reports a 26-participant study where layouts with crossings were fastest to interpret.
Purchase 1997 (understanding, abstract) and Ware 2002 (path finding, held, secondary) find the reverse.
Reading: the task decides. For path finding, avoid crossings or make them wide; for interpretation, crossings may not matter.

Thielsch 2019 reports g = 0.12 over 25 studies; Schlamann 2026 reports g = 0.29 over 31, and kept only studies whose aesthetic manipulation was confirmed by ratings. Both are heterogeneous, and neither is about figures.
(not verified: that later work contests the link from beauty to usability; it was named in a summary and not retrieved.)
Harrison 2015 shows appeal forms in 500 ms and rests largely on colourfulness and visual complexity. That predicts liking, not understanding.
Song 2025 contradicts itself on image-style bars.

Theory favours left to right; Figl 2015 found readers adapt to other directions with 44 participants. A convention, not a result.

Das 2024: image form beats text under token limits. Zhong 2024: text beats image on most structure questions, except cycle checks.
Zou 2024 compares different models. No same-model comparison of an authored diagram as source and as image was found; Zhong 2024 and Das 2024 make it for generated graphs.
Reading: advice that "models read source better" is untested for authored diagrams.